arXiv subjects

Dongdong She

Publications and source records attributed to Dongdong She.

At least 19 recordsLinked to original sources

arXiv2026

Safe to Check, Unsafe to Use: Relinking at the Compression Boundary of LLM Agents

Summarization-based prompt compression is increasingly used by LLM agents to shorten long, distributed contexts, but it shifts the security boundary: filters inspect the pre-compression prompt while the backend acts on a newly generated compressed context. We identify relinking, a compression-boundary vulnerability where the compressor behaves as a confused deputy, summarizing distributed, locally benign fragments into a complete malicious instruction. Unlike prompt injection, relinking need not place an explicitly malicious payload in the source context. We show that relinking arises from summarization itself: attention makes separated fragments jointly available, pre-training makes compatible fragments plausible to connect, and post-training favors compact backend-actionable summaries. We formalize the attacker-induced form as adversarial relinking and present Relink, an automated DSL-based tool that splits malicious payloads into benign fragments while keeping the complete payload absent before compression. Across four long-context agent benchmarks, Relink achieves 86.9% Relink Rate and Backend Action Rate versus 17.0% for clean-split controls. Existing defenses fail to reliably capture adversarial relinking; our KBRA defense reduces residual Backend Action Rate to 0.0%.

cs.CR↗

arXiv2026

Machine learning inference of fission yields from gamma spectroscopy for very low-yield nuclear test verification

Very low-yield nuclear tests pose a major verification challenge for the zero-yield standard of the Comprehensive Nuclear-Test-Ban Treaty (CTBT). The zero-yield standard prohibits any explosive experiment that produces a self-sustaining fission chain reaction while allowing subcritical experiments. Previous research shows that on-site gamma spectroscopy of post-test debris provides useful insight into the criticality level, although it remains heavily dependent on knowledge of certain experimental settings. Here, we adopt a new approach whereby machine learning models are trained on simulated gamma spectroscopy data to infer the fission yield of a nuclear very low-yield test. Using high-fidelity 3D Monte Carlo particle transport simulations, we generated gamma spectra measured outside containment vessels after very low-yield tests for 66 million representative scenarios. From these spectra, we extracted 82 fission-product-to-plutonium-239 peak ratios, then trained ML models for two tasks: (1) binary classification of whether a test exceeded a chosen yield threshold, and (2) regression to estimate the actual yield. We find that XGBoost performs best on the classification task across the most policy-relevant yield range. The classifier achieves high accuracy even for yields near the chosen threshold (e.g., >95% for yields +-100 g around a threshold at 1 kg TNT), and the regressor presents a mean absolute relative error of 12.4% for measurements taken a month to a year after the test. These results demonstrate that using machine learning to infer the yield of a past very low-yield nuclear test from gamma spectroscopy data is feasible and accurate. This approach can support efforts to establish a robust verification protocol for the zero-yield standard and could pave the way for a future yield threshold-based verification regime that is both technically feasible and politically viable.

physics.ins-det↗

arXiv2026

Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents

LLM agents convert model outputs into consequential actions, including communications, code changes, and financial transactions. Developers often trust evidence such as test results and execution logs. We identify a response path integrity gap in Bring Your Own Key configurations used by roughly 88 percent of mainstream agents. Because traffic passes through a user-authorized relay, the relay can modify plaintext LLM responses after alignment but before execution without breaking encryption. A minimal attack rewrites one execution bearing field and regenerates the remaining response using the user key while preserving the model style. Experiments reveal false green verification, where malicious code modifications pass public tests while silently defeating security checks. On APPS, 99.7 percent of publicly passing solutions retained downgraded behavior without developer-visible warnings. Tests on SWE bench, AgentDojo, and ASB across five frontier models show that single-field rewriting can redirect agents while preserving apparent task completion. We propose sign-c, a server-side scheme that authenticates execution bearing fields and outgoing queries. A local shim verifies them before action, while encryption protects confidentiality. The defense rejected all tampered responses with zero false rejections and only 0.0167 percent latency overhead.

cs.CR↗

arXiv2026

RandSet: Randomized Corpus Reduction for Fuzzing Seed Scheduling

Seed explosion is a fundamental problem in fuzzing seed scheduling, where a fuzzer maintains a huge corpus and fails to choose promising seeds. Existing works focus on seed prioritization but still suffer from seed explosion since corpus size remains huge. We tackle this from a new perspective: corpus reduction, i.e., computing a seed corpus subset. However, corpus reduction could lead to poor seed diversity and large runtime overhead. Prior techniques like cull_queue, AFL-Cmin, and MinSet suffer from poor diversity or prohibitive overhead, making them unsuitable for high-frequency seed scheduling. We propose RandSet, a novel randomized corpus reduction technique that reduces corpus size and yields diverse seed selection simultaneously with minimal overhead. Our key insight is introducing randomness into corpus reduction to enjoy two benefits of a randomized algorithm: randomized output (diverse seed selection) and low runtime cost. Specifically, we formulate corpus reduction as a set cover problem and compute a randomized subset covering all features of the entire corpus. We then schedule seeds from this small, randomized subset rather than the entire corpus, effectively mitigating seed explosion. We implement RandSet on three popular fuzzers: AFL++, LibAFL, and Centipede, and evaluate it on standalone programs, FuzzBench, and Magma. Results show RandSet achieves significantly more diverse seed selection than other reduction techniques, with average subset ratios of 4.03% and 5.99% on standalone and FuzzBench programs. RandSet achieves a 16.58% coverage gain on standalone programs and up to 3.57% on FuzzBench in AFL++, triggers up to 7 more ground-truth bugs than the state-of-the-art on Magma, while introducing only 1.17%-3.93% overhead.

cs.SE↗

arXiv2026

On Protecting Agentic Systems' Intellectual Property via Watermarking

The evolution of Large Language Models (LLMs) into agentic systems that perform autonomous reasoning and tool use has created significant intellectual property (IP) value. We demonstrate that these systems are highly vulnerable to imitation attacks, where adversaries steal proprietary capabilities by training imitation models on victim outputs. Crucially, existing LLM watermarking techniques fail in this domain because real-world agentic systems often operate as grey boxes, concealing the internal reasoning traces required for verification. This paper presents AGENTWM, the first watermarking framework designed specifically for agentic models. AGENTWM exploits the semantic equivalence of action sequences, injecting watermarks by subtly biasing the distribution of functionally identical tool execution paths. This mechanism allows AGENTWM to embed verifiable signals directly into the visible action trajectory while remaining indistinguishable to users. We develop an automated pipeline to generate robust watermark schemes and a rigorous statistical hypothesis testing procedure for verification. Extensive evaluations across three complex domains demonstrate that AGENTWM achieves high detection accuracy with negligible impact on agent performance. Our results confirm that AGENTWM effectively protects agentic IP against adaptive adversaries, who cannot remove the watermarks without severely degrading the stolen model's utility.

cs.AI↗

arXiv2026

From Similarity to Vulnerability: Key Collision Attack on LLM Semantic Caching

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86\% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

cs.CR↗

arXiv2025

When Compression Becomes an Attack Surface: Black-Box Attacks on Prompt-Compressed LLM Agents

Prompt compression is increasingly deployed in LLM agents to reduce latency and cost, but it also determines what the backend LLM ultimately sees. We show that, when trusted and untrusted inputs are compressed under a shared budget, this lossy transformation creates a new attack surface: by perturbing only untrusted inputs before compression, an adversary can cause the compressor to discard task-critical evidence or safety guardrails before inference. Unlike prompt injection, jailbreaks, or RAG poisoning, the attack target is the compressor rather than the backend LLM; the perturbation need not encode a meaningful instruction or survive compression. We formalize this vulnerability as adversarial information loss (AIL), the excess downstream distortion caused by adversarially steering a lossy compressor beyond benign compression alone. To exploit AIL, we present COMA, a transfer-based black-box attack that optimizes pre-compression perturbations using attacker-side surrogate compressors and backend LLMs. Across three tasks and six compressors, COMA achieves 0.71 average ASR, versus 0.21 for the strongest baseline, and transfers to two real-world agent case studies.

cs.CR↗

arXiv2025

QueryIPI: Query-agnostic Indirect Prompt Injection on Coding Agents

Modern coding agents integrated into IDEs orchestrate powerful tools and high-privilege system access, creating a high-stakes attack surface. Prior work on Indirect Prompt Injection (IPI) is mainly query-specific, requiring particular user queries as triggers and leading to poor generalizability. We propose query-agnostic IPI, a new attack paradigm that reliably executes malicious payloads under arbitrary user queries. Our key insight is that malicious payloads should leverage the invariant prompt context (i.e., system prompt and tool descriptions) rather than variant user queries. We present QueryIPI, an automated framework that uses tool descriptions as optimizable payloads and refines them via iterative, prompt-based blackbox optimization. QueryIPI leverages system invariants for initial seed generation aligned with agent conventions, and iterative reflection to resolve instruction-following failures and safety refusals. Experiments on five simulated agents show that QueryIPI achieves up to 87% success rate, outperforming the best baseline (50%). Crucially, generated malicious descriptions transfer to real-world coding agents, highlighting a practical security risk.

cs.CR↗

arXiv2025

Red-Teaming Coding Agents from a Tool-Invocation Perspective: An Empirical Security Assessment

Coding agents powered by large language models are becoming central modules of modern IDEs, helping users perform complex tasks by invoking tools. While powerful, tool invocation opens a substantial attack surface. Prior work has demonstrated attacks against general-purpose and domain-specific agents, but none have focused on the security risks of tool invocation in coding agents. To fill this gap, we conduct the first systematic red-teaming of six popular real-world coding agents: Cursor, Claude Code, Copilot, Windsurf, Cline, and Trae. Our red-teaming proceeds in two phases. In Phase 1, we perform prompt leakage reconnaissance to recover system prompts. We discover a general vulnerability, ToolLeak, which allows malicious prompt exfiltration through benign argument retrieval during tool invocation. In Phase 2, we hijack the agent's tool-invocation behavior using a novel two-channel prompt injection in the tool description and return values, achieving remote code execution (RCE). We adaptively construct payloads using security information leaked in Phase 1. In emulation across five backends, our method outperforms baselines on Claude-Sonnet-4, Claude-Sonnet-4.5, Grok-4, and GPT-5. On real agents, our approach succeeds on 19 of 25 agent-LLM pairs, achieving leakage on every agent using Claude and Grok backends. For tool-invocation hijacking, we obtain RCE on every tested agent-LLM pair, with our two-channel method delivering the highest success rate. We provide case studies on Cursor and Claude Code, analyze security guardrails of external and built-in tools, and conclude with practical defense recommendations.

cs.CR↗

arXiv2025

ZTaint-Havoc: From Havoc Mode to Zero-Execution Fuzzing-Driven Taint Inference

Fuzzing is a widely used technique for discovering software vulnerabilities, but identifying hot bytes that influence program behavior remains challenging. Traditional taint analysis can track such bytes white-box, but suffers from scalability issue. Fuzzing-Driven Taint Inference (FTI) offers a black-box alternative, yet typically incurs significant runtime overhead due to extra program executions. We observe that the commonly used havoc mutation scheme in fuzzing can be adapted for lightweight FTI with zero extra executions. We present a computational model of havoc mode, demonstrating that it can perform FTI while generating new test cases. Building on this, we propose ZTaint-Havoc, a novel, efficient FTI with minimal overhead (3.84% on UniBench, 12.58% on FuzzBench). We further design an effective mutation algorithm utilizing the identified hot bytes. Our comprehensive evaluation shows that ZTaint-Havoc, implemented in AFL++, improves edge coverage by up to 33.71% on FuzzBench and 51.12% on UniBench over vanilla AFL++, with average gains of 2.97% and 6.12% in 24-hour fuzzing campaigns.

cs.CR↗

arXiv2025

Empc: Effective Path Prioritization for Symbolic Execution with Path Cover

Symbolic execution is a powerful program analysis technique that can formally reason the correctness of program behaviors and detect software bugs. It can systematically explore the execution paths of the tested program. But it suffers from an inherent limitation: path explosion. Path explosion occurs when symbolic execution encounters an overwhelming number (exponential to the program size) of paths that need to be symbolically reasoned. It severely impacts the scalability and performance of symbolic execution. To tackle this problem, previous works leverage various heuristics to prioritize paths for symbolic execution. They rank the exponential number of paths using static rules or heuristics and explore the paths with the highest rank. However, in practice, these works often fail to generalize to diverse programs. In this work, we propose a novel and effective path prioritization technique with path cover, named Empc. Our key insight is that not all paths need to be symbolically reasoned. Unlike traditional path prioritization, our approach leverages a small subset of paths as a minimum path cover (MPC) that can cover all code regions of the tested programs. To encourage diversity in path prioritization, we compute multiple MPCs. We then guide the search for symbolic execution on the small number of paths inside multiple MPCs rather than the exponential number of paths. We implement our technique Empc based on KLEE. We conduct a comprehensive evaluation of Empc to investigate its performance in code coverage, bug findings, and runtime overhead. The evaluation shows that Empc can cover 19.6% more basic blocks than KLEE's best search strategy and 24.4% more lines compared to the state-of-the-art work cgs. Empc also finds 24 more security violations than KLEE's best search strategy. Meanwhile, Empc can significantly reduce the memory usage of KLEE by up to 93.5%.

cs.CR↗

arXiv2025

RouthSearch: Inferring PID Parameter Specification for Flight Control Program by Coordinate Search

Flight control programs use PID control modules with user-configurable Proportional (P), Integral (I), and Derivative (D) parameters to manage UAV flying behaviors. Users can adjust these PID parameters during flight. However, flight control programs lack sufficient safety checks on user-provided PID parameters, leading to a severe UAV vulnerability - the input validation bug. This occurs when a user misconfigures PID parameters, causing dangerous states like deviation from the expected path, loss of control, or crash. Prior works use random testing like fuzzing, but these are not effective in the three-dimensional search space of PID parameters. The expensive dynamic execution of UAV tests further hinders random testing performance. We address PID parameter misconfiguration by combining the Routh-Hurwitz stability criterion with coordinate search, introducing RouthSearch. Instead of ad-hoc identification, RouthSearch principledly determines valid ranges for three-dimensional PID parameters. We first leverage the Routh-Hurwitz Criterion to identify a theoretical PID parameter boundary, then refine it using efficient coordinate search. The determined valid range can filter misconfigured PID parameters from users during flight and help discover logical bugs in flight control programs. We evaluated RouthSearch across eight flight modes in PX4 and Ardupilot. Results show RouthSearch determines valid ranges with 92.0% accuracy compared to ground truth. RouthSearch discovers 3,853 PID misconfigurations within 48 hours, while the STOA work PGFuzz discovers only 449 sets, significantly outperforming prior works by 8.58 times. Our method also helped detect three bugs in ArduPilot and PX4.

cs.SE↗

arXiv2025

Extracting Formal Specifications from Documents Using LLMs for Automated Testing

Automated testing plays a crucial role in ensuring software security. It heavily relies on formal specifications to validate the correctness of the system behavior. However, the main approach to defining these formal specifications is through manual analysis of software documents, which requires a significant amount of engineering effort from experienced researchers and engineers. Meanwhile, system update further increases the human labor cost to maintain a corresponding formal specification, making the manual analysis approach a time-consuming and error-prone task. Recent advances in Large Language Models (LLMs) have demonstrated promising capabilities in natural language understanding. Yet, the feasibility of using LLMs to automate the extraction of formal specifications from software documents remains unexplored. We conduct an empirical study by constructing a comprehensive dataset comprising 603 specifications from 37 documents across three representative open-source software. We then evaluate the most recent LLMs' capabilities in extracting formal specifications from documents in an end-to-end fashion, including GPT-4o, Claude, and Llama. Our study demonstrates the application of LLMs in formal specification extraction tasks while identifying two major limitations: specification oversimplification and specification fabrication. We attribute these deficiencies to the LLMs' inherent limitations in processing and expressive capabilities, as well as their tendency to fabricate fictional information. Inspired by human cognitive processes, we propose a two-stage method, annotation-then-conversion, to address these challenges. Our method demonstrates significant improvements over the end-to-end method, with a 29.2% increase in the number of correctly extracted specifications and a 14.0% improvement in average accuracy. In particular, our best-performing LLM achieves an accuracy of 71.6%.

cs.SE↗

arXiv2024

Comment on Revisiting Neural Program Smoothing for Fuzzing

MLFuzz, a work accepted at ACM FSE 2023, revisits the performance of a machine learning-based fuzzer, NEUZZ. We demonstrate that its main conclusion is entirely wrong due to several fatal bugs in the implementation and wrong evaluation setups, including an initialization bug in persistent mode, a program crash, an error in training dataset collection, and a mistake in fuzzing result collection. Additionally, MLFuzz uses noisy training datasets without sufficient data cleaning and preprocessing, which contributes to a drastic performance drop in NEUZZ. We address these issues and provide a corrected implementation and evaluation setup, showing that NEUZZ consistently performs well over AFL on the FuzzBench dataset. Finally, we reflect on the evaluation methods used in MLFuzz and offer practical advice on fair and scientific fuzzing evaluations.

cs.CR↗

arXiv2024

FOX: Coverage-guided Fuzzing as Online Stochastic Control

Fuzzing is an effective technique for discovering software vulnerabilities by generating random test inputs and executing them against the target program. However, fuzzing large and complex programs remains challenging due to difficulties in uncovering deeply hidden vulnerabilities. This paper addresses the limitations of existing coverage-guided fuzzers, focusing on the scheduler and mutator components. Existing schedulers suffer from information sparsity and the inability to handle fine-grained feedback metrics. The mutators are agnostic of target program branches, leading to wasted computation and slower coverage exploration. To overcome these issues, we propose an end-to-end online stochastic control formulation for coverage-guided fuzzing. Our approach incorporates a novel scheduler and custom mutator that can adapt to branch logic, maximizing aggregate edge coverage achieved over multiple stages. The scheduler utilizes fine-grained branch distance measures to identify frontier branches, where new coverage is likely to be achieved. The mutator leverages branch distance information to perform efficient and targeted seed mutations, leading to robust progress with minimal overhead. We present FOX, a proof-of-concept implementation of our control-theoretic approach, and compare it to industry-standard coverage-guided fuzzers. 6 CPU-years of extensive evaluations on the FuzzBench dataset and complex real-world programs (a total of 38 test programs) demonstrate that FOX outperforms existing state-of-the-art fuzzers, achieving average coverage improvements up to 26.45% in real-world standalone programs and 6.59% in FuzzBench programs over the state-of-the-art AFL++. In addition, it uncovers 20 unique bugs in popular real-world applications including eight that are previously unknown, showcasing real-world security impact.

cs.CR↗

arXiv2022

NeuDep: Neural Binary Memory Dependence Analysis

Determining whether multiple instructions can access the same memory location is a critical task in binary analysis. It is challenging as statically computing precise alias information is undecidable in theory. The problem aggravates at the binary level due to the presence of compiler optimizations and the absence of symbols and types. Existing approaches either produce significant spurious dependencies due to conservative analysis or scale poorly to complex binaries. We present a new machine-learning-based approach to predict memory dependencies by exploiting the model's learned knowledge about how binary programs execute. Our approach features (i) a self-supervised procedure that pretrains a neural net to reason over binary code and its dynamic value flows through memory addresses, followed by (ii) supervised finetuning to infer the memory dependencies statically. To facilitate efficient learning, we develop dedicated neural architectures to encode the heterogeneous inputs (i.e., code, data values, and memory addresses from traces) with specific modules and fuse them with a composition learning strategy. We implement our approach in NeuDep and evaluate it on 41 popular software projects compiled by 2 compilers, 4 optimizations, and 4 obfuscation passes. We demonstrate that NeuDep is more precise (1.5x) and faster (3.5x) than the current state-of-the-art. Extensive probing studies on security-critical reverse engineering tasks suggest that NeuDep understands memory access patterns, learns function signatures, and is able to match indirect calls. All these tasks either assist or benefit from inferring memory dependencies. Notably, NeuDep also outperforms the current state-of-the-art on these tasks.

cs.CR↗

arXiv2022

$MC^2$: Rigorous and Efficient Directed Greybox Fuzzing

Directed greybox fuzzing is a popular technique for targeted software testing that seeks to find inputs that reach a set of target sites in a program. Most existing directed greybox fuzzers do not provide any theoretical analysis of their performance or optimality. In this paper, we introduce a complexity-theoretic framework to pose directed greybox fuzzing as a oracle-guided search problem where some feedback about the input space (e.g., how close an input is to the target sites) is received by querying an oracle. Our framework assumes that each oracle query can return arbitrary content with a large but constant amount of information. Therefore, we use the number of oracle queries required by a fuzzing algorithm to find a target-reaching input as the performance metric. Using our framework, we design a randomized directed greybox fuzzing algorithm that makes a logarithmic (wrt. the number of all possible inputs) number of queries in expectation to find a target-reaching input. We further prove that the number of oracle queries required by our algorithm is optimal, i.e., no fuzzing algorithm can improve (i.e., minimize) the query count by more than a constant factor. We implement our approach in MC$^2$ and outperform state-of-the-art directed greybox fuzzers on challenging benchmarks (Magma and Fuzzer Test Suite) by up to two orders of magnitude (i.e., $134\times$) on average. MC$^2$ also found 15 previously undiscovered bugs that other state-of-the-art directed greybox fuzzers failed to find.

cs.CR↗

arXiv2022

Effective Seed Scheduling for Fuzzing with Graph Centrality Analysis

Seed scheduling, the order in which seeds are selected, can greatly affect the performance of a fuzzer. Existing approaches schedule seeds based on their historical mutation data, but ignore the structure of the underlying Control Flow Graph (CFG). Examining the CFG can help seed scheduling by revealing the potential edge coverage gain from mutating a seed. An ideal strategy will schedule seeds based on a count of all reachable and feasible edges from a seed through mutations, but computing feasibility along all edges is prohibitively expensive. Therefore, a seed scheduling strategy must approximate this count. We observe that an approximate count should have 3 properties -- (i) it should increase if there are more edges reachable from a seed; (ii) it should decrease if mutation history information suggests an edge is hard to reach or is located far away from currently visited edges; and (iii) it should be efficient to compute over large CFGs. We observe that centrality measures from graph analysis naturally provide these three properties and therefore can efficiently approximate the likelihood of reaching unvisited edges by mutating a seed. We therefore build a graph called the edge horizon graph that connects seeds to their closest unvisited nodes and compute the seed node's centrality to measure the potential edge coverage gain from mutating a seed. We implement our approach in K-scheduler and compare with many popular seed scheduling strategies. We find that K-scheduler increases feature coverage by 25.89% compared to Entropic and edge coverage by 4.21% compared to the next-best AFL-based seed scheduler, in arithmetic mean on 12 Google FuzzBench programs. It also finds 3 more previously-unknown bugs than the next-best AFL-based seed scheduler.

cs.CR↗